MACHINE LEARNING DRIVEN CROP YIELD PREDICTION AND AGRICULTURE DECISION SUPPORT SYSTEM USING HISTORICAL WEATHER, SOIL AND SATELLITE DATA FOR INDIAN FARMERS

Accurate, farmer-usable crop yield forecasting remains an open problem in Indian agriculture because rainfall, temperature, humidity, and soil fertility vary sharply across districts, while most published machine learning (ML) solutions are trained and validated at a coarse state level using only two or three input variables.This paper proposes a district-level, multi-source data fusion framework for crop yield prediction and agricultural decision support, combining historical weather records, soil health card parameters (nitrogen, phosphorus, potassium, pH, organic carbon), satellite-derived vegetation indices (NDVI), and a novel district-similarity network feature.Random Forest and XGBoost regressors are proposed as the core predictive engines, augmented with a SHAP-based explainability layer and a confidence-scoring mechanism reflecting historical data density for a given district-crop pair.The trained model is re-used to rank candidate crops for a farmer's specific soil and weather profile, producing an actionable crop recommendation.The pipeline is implemented as a deployed system with Flask REST APIs, a MySQL persistence layer, a Streamlit dashboard, and a ruleassisted chatbot.A structured survey of twenty-nine studies identifies four recurring gaps in existing work, and the proposed architecture is designed to close all four simultaneously.